Edge Computing Resource Management for Cross-Camera Video Analytics: Workload and Model Adaptation

Multi-camera systems are now widely employed across numerous domains. The exponential growth of deep learning has simplified the implementation of advanced video analytics applications. While current systems strive to enhance live video analytics from several aspects, they overlook the potential degradation in performance resulting from dynamic content changes, such as the variation in the quantity and types of objects of interest over a period and across different cameras. The authors introduce Workload and Model Adaptation (WMA), a two-stage resource allocation strategy for a three-tiered, cross-camera video analytics system. This system not only supports model fine-tuning but also ensures workload balance. Notably, both the system architecture and control workflow fully comply with the IEEE 1935 edge standard. This paper delves into the GPU utilization performance of a vehicle re-identification application and examines the workload dynamics spanning multiple cameras. Furthermore, the challenges related to multi-process execution are explored. The system is evaluated using a commonly employed dataset and a popular open-source project. The results demonstrate that the proposed design surpasses the baseline and enhances the overall throughput and latency across cameras within the system.